Skip to main content

Capacity assessment and calibration

Attempt independently before calibration. Score 0 absent/incorrect, 1 prompted/incomplete, 2 independent correct evidence with limits, 3 justified fresh transfer. Every category requires at least 2; a passing arithmetic harness does not replace representative measurement or design review.

Required evidence​

  • SD1 — Units: supply an assumption ledger with units, workload mix, source/range, downstream amplification and measurement plan. Separate policy headroom from measured sustainable capacity.
  • SD2 — Sizing: derive minimum replica count and verify feasibility under the declared failures; state balancing, linearity and failure-domain assumptions.
  • SD3 — Queue: hand-predict burst/outage/recovery traces, implement the archive contract and add three independent fixtures. Explain conservation and the chosen slot ordering.
  • SD4 — Sensitivity: compare 10x traffic, 100x objects, tighter freshness and a hot key. Identify changed mechanisms, failure placement and the next experiment.
  • SD5 — Measurement: run a representative small implementation under recorded load and one dependency failure. Compare prediction with observation, record discrepancies and avoid claiming measured p95 from the bulk model.
  • SD6 — Design defense: preserve original API/data model, two deep dives, write/read traces, failure matrix, alternatives and review feedback. Revise a decision from evidence and identify remaining uncertainty.

Calibration — after attempting​

For demand 900, capacity 200, utilization policy 3/4 and one failed replica, N=7; N=6 supplies only 750. At capacity 150, N=9 supplies 900 after one failure. Zero demand returns the failure-count parameter under the lab's arithmetic rule, not a recommendation to operate that many idle machines.

For [6,0,0] arrivals, [2,2,2] service and limit 4, two units drop, four complete, final backlog is zero. Spread arrivals [2,2,2] complete all six. No p95 follows from either trace. A hot key can defeat fleet-level balance; region loss must count actual lost replicas and shared dependencies. Increasing the queue limit can reduce immediate drops while violating completion/freshness deadlines.

If an experiment disagrees with the model, inspect measurement coverage and assumptions before blaming the implementation. A passing conservation check establishes accounting, not a service-level objective.

Delayed transfer​

After at least 48 hours, independently solve a new workload with different units and a correlated failure. Preserve errors, repairs and assistance disclosures. Re-test the weakest category before advancing; author reference results are not learner measurements.